| | --- |
| | license: apache-2.0 |
| | tags: |
| | - krystalmind |
| | - pdssm |
| | - pharma-crm |
| | - state-space-model |
| | --- |
| | |
| | # KrystalMind Synapse (PD-SSM) |
| |
|
| | Pharma-Domain State Space Model for real-time entity divergence detection. |
| |
|
| | ## Architecture |
| | - **41.5M params**: PD-SSM 38.4M + concept_projector 2.0M + heads 1.1M |
| | - **6 SSM layers**: Pharma-domain diagonal SSM with complex hidden state |
| | - **Graph-Guided SSM (GG-SSM)**: 31,494 additional params for neighbor-aware selector blending |
| | - **3 semantic tiers**: Tactical (L1-2), Relational (L3-4), Strategic (L5-6) |
| | - **STATE_DIM**: 128 (complex: re + im per layer) |
| | |
| | ## Files |
| | - `phase2_ggssm_routing.km` — Production baked model with GG-SSM + routing head (58 MB) |
| | - `phase2_ggssm_routing.safetensors` — SafeTensors format (156 MB) |
| | - `routing_head_best.pt` — MLP routing classifier (1.5 MB) |
| | - `checkpoints/phase1_production.pt` — Phase 1 best checkpoint, step 1500 (460 MB) |
| |
|
| | ## Training |
| | - **Phase 1**: 35,700 sequences (34 entity types), avg FSA=93%, tier_corr=0.937 |
| | - **Phase 2**: GG-SSM graph messages, val_loss=0.66, 22/22 tests passing |
| | - **Latency**: 1-2ms per entity scan |
| |
|
